Comment by titanix88

Looks like the author have not used software pipelining compiler directives with the kernel loops. AMD AIE architecture has 5 cycle load/store latency and 7 cycle FP unit latency. With software pipelining, they could have 5-10x speed up for long loops.